Who Becomes An AI Market Surveillance Officer, And How Do You Hire One?

Why Your Shortlist Keeps Coming Back Empty

You wrote a vacancy notice against Article 70(3) and it asked for one person who understands AI technologies, data and computing, personal data protection, cybersecurity, fundamental rights, and health and safety risks ¹. Forty applications arrived. Most were machine learning graduates with no casework, a few were career inspectors who declared no AI knowledge, and none were both.

That is not a pipeline failure. The competence list in Article 70(3) describes what the authority must hold, not what one hire must carry, and the annual reassessment duty in the same paragraph reads as a portfolio obligation rather than a person specification. Read it as a team brief and the search becomes solvable: you are hiring one profile at a time against a competence map, and you should know which square you are filling before the notice goes out.

The practical consequence is that the first hire should be the one who can run a file end to end. An investigation that stalls because nobody can compel documentation, scope a request proportionately, or write an adverse finding in language that survives an appeal is worse than no investigation, because it burns the authority's credibility with the exact operators it will need to inspect again next year.

Which Traits Separate An Inspector Who Closes Files From One Who Only Opens Them?

The tell is what a candidate does with a claim they cannot check. Ask about a time a company told them something plausible that they had no way to verify. The real inspector describes what they asked for next, what they did when it was refused, and how the file recorded the gap. The performed one describes how they could tell the company was lying, which is a claim about intuition and is not evidence of anything.

Four traits matter more than the rest. First, proportionality instinct: a request for every training artifact a deployer holds is unenforceable in practice and will be litigated into nothing, while a narrow request for the logs covering one deployment window usually arrives. Second, comfort with partial technical understanding, because an inspector will never match a vendor's engineers and the useful skill is knowing which three questions expose whether the vendor's own testing was real. Third, written precision, since the finding is the product. Fourth, the willingness to record an inconclusive result rather than round it up to a violation.

The counter-tell is the candidate who talks about catching bad actors. Market surveillance is mostly conformity checking on operators who believe they are compliant, and the frequent outcome is a documentation gap, a corrective action, and a follow-up. Someone who arrives hunting fraud will over-escalate the ordinary case and exhaust the authority's legal capacity on files that should have closed with a letter.

One more, easy to miss in an interview panel: ask what they would do if their own draft finding turned out to rest on a misread specification after it had gone to the operator. You are listening for a correction procedure, not for contrition.

Which Backgrounds Actually Produce This Person?

The productive pools are inspectorates that already handle technical files under a conformity regime. National product safety authorities, medical device competent authorities, telecoms and radio equipment regulators, metrology institutes, and the casework side of data protection authorities all train people to read a technical file, test a claim against a standard, and write a decision. Those staff already hold the competences in Article 70(3) that are hardest to teach, and the AI-specific parts are the additions ¹.

The unexpected pools are better than they look. Aviation and rail safety investigators are trained to reconstruct a failure from logs and to resist the first plausible cause, which is close to what an inspector does with an incident report on a high-risk system. Financial market surveillance analysts already read automated systems for behaviour rather than for intent. Clinical trial monitors audit data provenance against a protocol for a living. Public procurement officers who have run technical evaluations of software know exactly how vendor claims are written to be technically true.

Engineers are worth pursuing, but selectively. The ones who convert are those who have worked in evaluation, red teaming, model validation in a regulated bank, or safety engineering, because they already accept that their job is to find what is wrong and put it in writing. A research scientist moving to a regulator for the first time will usually struggle with the pace and the paperwork, and both of you will know within six months. If your authority also touches deployment-side obligations, the hiring logic overlaps with an AI Act enforcement officer and with the assessment work described in hiring a fundamental rights impact assessment lead.

How This Person Got Good At AI Without An AI Job

The strong candidates did not take a course. They used the systems adversarially in their own casework and kept notes on where the systems failed them. That produces a specific kind of literacy: they know what a model output looks like when it is confident and wrong, because they have been handed one and nearly filed it.

Ask what they used an assistant for in the last six months and what it got wrong. The answer you want is concrete. They drafted a request for information and found the assistant had invented a legal basis, so they now check every citation against the primary text before it leaves the file. They used a model to summarise a two hundred page technical file and discovered a whole annex had been compressed into one sentence that reversed its meaning, so they now spot-check summaries against the sections that carry the obligation. They asked a model to explain an evaluation metric and got a plausible definition that did not match the vendor's own documentation, which told them the vendor was using the term loosely.

Candidates who talk about productivity gains without a single failure story have either not used these systems on anything that mattered or have not been checking. For an inspector, the second is disqualifying. The habit that transfers is verification against a source outside the conversation, and it is visible in how someone works rather than in what they say about their tooling.

Where To Find Them, And How To Close Them

Recruit laterally inside the state before you advertise. The people you want are already civil servants in another inspectorate, which means secondment or internal mobility can move them in weeks rather than the months an external competition takes, and it carries their security clearance and their institutional standing with it.

Standing venues that actually exist and are worth working: the European Commission's AI Office and the AI Board's national contact points, the ICSMS and Safety Gate communities where market surveillance staff already coordinate, CEN-CENELEC JTC 21 standardisation working groups, national metrology and notified body networks, and the data protection conference circuit where enforcement staff gather. Trade unions and staff associations inside the existing authorities are an underused route to people who are ready to move.

Closing is not a compensation exercise, because you will usually lose that comparison and both sides know it. The offer that works is scope and access. Name the sectors this person will own, say how many other people will be on the team by a specific date, and be honest about the current caseload. Give them a title with inspector or officer in it that reads as authority to an operator's counsel, because that is a real working asset.

Two things close a hesitant candidate. First, a written commitment to training time on real systems, including budget for access to the tools they will be inspecting. Second, a named senior person who will back an adverse finding. Candidates leaving other inspectorates have almost always been undercut once, and they will ask. Answer that question honestly, including where the authority has folded, because they will find out and a discovered omission costs you the hire in the first year instead of at the offer stage.

What To Pay, And Where This Job Sits

There is no market rate for this title yet. The category is forming and the public postings are still few, so the honest answer is which band you hire against. This role sits with senior technical inspector and senior enforcement casework grades in the same authority, and it belongs at the top of that band rather than the middle, because the competence list in Article 70(3) is broader than what those grades usually carry ¹.

Say out loud that the band is a proxy. It is derived from the grades the realistic candidate already holds, not from any survey of this post, and any point estimate you read for the title is somebody's guess. Where a national scheme allows a technical or scarce-skills supplement, this is the role it exists for, and the supplement is the honest place to put the difference rather than a grade you cannot defend to your own HR function.

What a grade cannot fix is the comparison the candidate is running privately. One vendor analysis, drawn from roughly a billion job advertisements across every occupation, put the average wage premium for roles demanding AI skills at around 62 percent ². That is an occupation-wide average and it describes nothing about this post; it is quoted here only to name the size of the pull the private sector exerts on your shortlist, and it is not a basis for a band. A public authority will not match it, and pretending otherwise in an interview costs credibility. What a state offers instead is the work itself, statutory backing, and access no private employer can provide. Comparable pricing pressure shows up in adjacent commercial roles such as an AI product counsel.

On location, treat physical presence as a requirement with narrow exceptions rather than a preference. Inspection work involves confidential technical files, on-site visits to operators, and evidence handling that most authorities restrict to secure premises. A hybrid pattern of two or three fixed office days is common and workable. Fully remote is realistic only for desk-based conformity review, and if you advertise it that way you should say which parts of the job it excludes. Cross-border cooperation under the Act also means real travel, which some candidates want and others will decline over, so put the expected days in the notice.

One caveat worth stating to a hiring committee: national implementation differs, deadlines and designated authorities vary by member state, and the position of this role in a civil service grading scheme is a question for your own HR and legal function. Check the national implementing measures and take advice from counsel before you fix a grade or publish a competence profile.

Read the evidence >